Version 1.3.5: Large Turbo transcribes about five times faster — and Harvestry is far better tested.
Someone trying Harvestry wrote in recently: each 90-minute lecture took over half an hour to transcribe with Large Turbo on a MacBook Air. Was it even using the GPU? The heavy lifting had been happening all along on the Neural Engine, the part of Apple silicon built for exactly this. The real problem was ours: what we'd shipped as Large Turbo wasn't the turbo model at all. 1.3.5 fixes that, and it's built on a much bigger set of automated tests.
The real Large Turbo
What Harvestry downloaded as Large Turbo was OpenAI's full Whisper Large v3: 3.2 GB, and every word it wrote passed through 32 stages. The real Large v3 Turbo does that part in 4, which is where the speed comes from, and it's half the size. 1.3.5 switches to it.
Settings → Tools now shows the real download sizes, too: about 490 MB for Small, 1.5 GB for Medium and 1.6 GB for Large Turbo. Small and Medium had been listed at about half their real size, and the old Large Turbo said ~800 MB but took 3.2 GB.
Timing transcription on its own, with Harvestry's settings, a 58-minute webinar took 29 minutes on our M2 Max before and 5.4 minutes now. About 96% of the words match; most differences are names, medical terms and numbers, and each model gets some right that the other misses. But the old model skipped a 30-second passage, and with the spoken language left on Auto-detect, the new one caught it in all three of our test runs. Screenshot capture runs alongside transcription and hasn't changed, so the whole job won't finish five times sooner.
As for the GPU: Activity Monitor shows the CPU and GPU but has no column for the Neural Engine, so the GPU looks idle and the work seems to be on the CPU. On our Mac, the old model alone kept about two CPU cores busy; the new one keeps about one.
A speed-up we didn't ship
Transcribing the audio in pieces, side by side, made the new model another 1.5 times faster. But it lost six stretches of 15–25 seconds of speech, and some pieces of an English lecture came back in German or Spanish. A transcript that's quietly missing sentences is worse than a slower one.
If you've downloaded Large Turbo
There's nothing to do. The first time you open 1.3.5, Harvestry downloads the new model in the background (about 1.6 GB) and keeps using the old one until it's done, then removes the old 3.2 GB one — or, if the old one is in use at that moment, the next time you open Harvestry. Harvestry now downloads one model at a time, so the other Download buttons in Settings → Tools wait until it's finished. Lectures you've made with Large Turbo won't ask to be processed again. And because the turbo model wasn't trained for translation, choosing Translate to English (Pro) still suggests switching to Medium.
Testing for a bigger Harvestry
Since it launched in June, Harvestry has gained translation, a processing queue, slideshows, PDF documents, Spotlight search and more, and every change can quietly break something somewhere else. That's why much of this release went into automated tests: small programs we can run after every change to check that Harvestry still does what it should.
Harvestry 1.3 had 157 of them; 1.3.5 has 590. Between them, they now put about nine in ten lines of the code that does the work (screenshots, downloads, the processing queue, export, the library) through their paces, up from about half in 1.3. For the windows and buttons themselves, 29 more tests open Harvestry and click through it the way you would, from the welcome screen to processing, Stop Processing and every delete confirmation. They run on a separate virtual Mac with its own throwaway library, so they never touch a real one.
One more test, which we run by hand, puts a real lecture through the real speech model and flags any stretch of 15 seconds or more where someone is talking but nothing was transcribed. On our 58-minute webinar, with the language on Auto-detect, it flags the old Large Turbo's missing 30 seconds, and the new model passes. Writing these tests turned up most of the fixes below.
Things that used to go wrong
Stop Processing (or Abort Processing in the lecture view) now stops consolidation too, the step that writes your consolidated notes, rather than leaving the queue “stopping” for minutes and then saving the notes into the lecture you stopped. If processing stops by itself, for example because a model download was interrupted, the lecture now goes back to Ready instead of staying stuck partway through. A downloaded video can no longer import as one of its temporary pieces instead of the finished file, and Stop Download no longer says “Download failed” or trips up an immediate retry.
Delete All and Delete Unprocessed now also remove the videos Harvestry downloaded from a URL, as deleting a single lecture already did. Videos you imported from your own folders are never touched. Deleted lectures no longer linger in the sidebar, and notes can no longer appear twice in the transcript. And the welcome screen now fits small screens: at 1024×768, Get Started and Don't show this again were off the bottom.
Who gets what
Everything in this release is in both Standard and Pro. Translate to English is part of Pro, so only Pro users (and anyone on the free trial) will see the suggestion to use Medium.
What's in v1.3.5
- Large Turbo transcribes about 5× faster.In our test, a 58-minute webinar went from 29 minutes to 5.4. It's OpenAI's real turbo model, at half the size.
- Automatic switch-over.If you've downloaded Large Turbo, the new model arrives in the background, and you get 1.6 GB of disk space back.
- Real download sizes.Settings → Tools now shows what each model really takes.
- 590 automated tests.Up from 157, plus 29 that click through the app itself.
- Stop means stop.Including consolidation.
- Steadier downloads.No temporary pieces in place of the video, and no false “Download failed.”
- Cleaner deleting.Delete All and Delete Unprocessed remove the videos Harvestry downloaded from a URL — never your own files.
- Also fixed.The welcome screen on small screens, screenshots that vanished when iCloud Drive couldn't be written, a newly chosen export folder that sometimes didn't stick, a rare sidebar crash during iCloud sync, and files left behind by a failed ffmpeg install.